1. No category

Exploring Communities for Effective Location Prediction
Jun Pang
Yang Zhang
University of Luxembourg
University of Luxembourg
jun.pang@uni.lu
yang.zhang@uni.lu
ABSTRACT
Humans are social animals, they interact with different communities to conduct different activities. The literature has
shown that human mobility is constrained by their social relations. In this work, we investigate the social impact on a
user’s mobility from his communities in order to conduct location prediction effectively. Through analysis of a real-life
dataset, we demonstrate that (1) a user gets more influences
from his communities than from all his friends; (2) his mobility is influenced only by a small subset of his communities;
(3) influence from communities depends on social contexts.
We further exploit a SVM to predict a user’s future location
based on his community information. Experimental results
show that the model based on communities leads to more
effective predictions than the one based on friends.
Categories and Subject Descriptors
H.2.8 [Database Management]: Database Applications—
Data mining
1.
INTRODUCTION
Humans are social animals, everyone is a part of the society and receives influences from it. For example, our daily
behaviors such as what types of music to listen or where to go
for lunch are largely dependent on our social relations. Sociology has shown that we can categorize our social relations
(or friends) into communities, using different criteria and
considerations. In daily life, humans are engaged in various
social environments, and interact with different communities
depending on the environments. Therefore, for a specific behavior of a user, in most cases social influence comes from
one of his communities, but not from all his friends. For example, one listens to similar music as his close friends; and
he has lunch together with his colleagues on weekdays.
With the large amount of location (check-in) and social
relation data from location-based social networks (LBSNs)
available, studying human mobility and its connection with
social relationships become quantitatively possible. Understanding human mobility can lead to compelling applications
Permission to make digital or hard copies of part or all of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage, and that copies
bear this notice and the full citation on the first page. Copyrights for thirdparty components of this work must be honored. For all other uses, contact
the owner/author(s). Copyright is held by the author/owner(s).
WWW 2015 Companion, May 18–22, 2015, Florence, Italy.
ACM 978-1-4503-3473-0/15/05.
http://dx.doi.org/10.1145/2740908.2742720.
including location recommendation, urban planning, etc. In
this paper, we aim to study the impact from communities on
a user’s mobility and effectively predict his future locations
based on his community information.
2.
COMMUNITY AND FRIENDS
The LBSN dataset we use is collected by the authors of [1].
We focus on the check-in data from four metropolises in US
including New York, LA, Bay Area and Dallas. We apply
the widely used algorithm InfoMap to partition each user’s
friends into communities. If a user is engaged in a number of
communities, then he is considered an active society member. Therefore, his daily behaviors are largely dependent on
his social relations. To quantify a user u’s social diversity,
we introduce the notion of community entropy:
X
|c(u)| α
1
log
) .
(
coment(u) =
1−α
|f (u)|
c(u)∈C(u)
The community entropy is defined based on R´enyi entropy
where C(u) is the set of u’s communities, c(u) is one of u’s
community containing a subset of his friends, f (u) is the set
of all u’s friends, and α is the order of diversity (set to 10).
For a user u, we exploit single linkage clustering algorithm
to group his check-ins into his frequent movement areas fa(u)
(a set of his clusters’ central points). Then, the impact from
his communities on his mobility is defined as
X
1
im(u)c =
min{d (l,{fa(c(u))|∀c(u) ⊆ C(u)})},
|fa(u)|
l∈fa(u)
where d represents the distances between {fa(c(u))|∀c(u) ⊆
C(u)} and l. The community impact is the average of all
the shortest distances between a user’s frequent movement
areas and any movement areas from any communities of the
user. Moreover, the communities that are used when computing im(u)c are named u’s influential communities. We
define a user’s friends impact on his mobility in a similar
way. A smaller distance indicates more impact on mobility.
Therefore, if im(u)f > im(u)c , then u gets more impact on
his mobility from his communities than his friends.
600
400
# im(u i)c < im(ui)f
200
# im(u i)c > im(ui)f
0
# im(u i)c = im(ui)f
New York
L.A.
Bay Area
Dallas
Figure 1: The numbers of users who get more, less or equal
impact from their communities than their friends.
As shown in Fig. 1, in general more users get more impact from communities than friends in the four metropolises.
0.5
1
1.5
community entropy
600
1
10
New York
L.A.
Bay Area
(a) Temporal
Dallas
0
New York
L.A.
Bay Area
Dallas
(b) Spatial
Figure 4: The number of common and distinct influential
communities under temporal or spatial contexts.
In New York, Bay Area and Dallas, almost twice users get
more impact from communities than friends. Fig. 2 shows
that more diverse a user’s community (as described by community entropy), more probably he gets more influence from
his communities than friends.
3.
COMMUNITY AND MOBILITY
Next, we focus on how many communities can actually
influence a user’s mobility. We plot the distribution of the
number of user’s influential communities in Fig. 3. This distribution follows the power law, indicating that most of users
are influenced only by a small subset of their communities.
Different communities give influences under different social contexts. For instance, a user has lunch with his colleagues and spends time with his family near his home. To
study this, we proceed by considering temporal and spatial
contexts. For temporal context, we choose two time periods
including lunch (11am–1pm) and dinner (7pm–9pm) hours
on weekdays. We first find users’ frequent movement areas
during lunch and dinner respectively, and then find the influential communities for users w.r.t. these two time periods.
Fig. 4a’s result indicates that the influential communities
of users during lunch and dinner time are quite different.
This simply reflects the fact that the people that users have
lunch and dinner with are quite different. For spatial contexts, we pick two disjoint areas in each metropolis. Fig. 4b
shows again the number of common and distinct influential
communities are quite different, meaning that the influential
communities are constrained by spatial contexts as well.
4.
LOCATION PREDICTION
Following the above analysis, we continue to investigate
whether it is possible to use community information to effectively predict users’ locations, using machine learning classifiers. The question that we want to answer is: given a
user’s community information, whether he will check in at
a certain place in the future.
Features. The features we use in the machine learning
classifier are from three domains including community, time
and location. For the community-related features, we first
pick the closest community to the location, and then use six
features including the community’s distance to the location,
the community size, the number of its frequent movement
1.5
community entropy
2
1
friends
0.95 community
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0
0.5
1
1.5
community entropy
(a) New York
2
(b) Bay Area
Figure 5: Prediction accuracy
same communities
different communities
200
200
accuracy
accuracy
2
10
# of communities
1
friends
0.95 community
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0
0.5
1
Figure 3: Distribution of #
of influential communities
400
400
0
0
600
same communities
different communities
1
10
10 0
10
2
Figure 2: Users having more
impact from communities
2
10
precision
New York
L.A.
Bay Area
Dallas
New York
L.A.
Bay Area
Dallas
1
1
0.9
0.9
precision
10
# of users
ratio
3
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0
0.8
0.7
0.6
0.5
0
0.7
0.6
friends
community
0.2
0.8
0.4
0.6
recall
0.8
(a) New York
1
0.5
0
friends
community
0.2
0.4
0.6
recall
0.8
1
(b) Bay Area
Figure 6: Precision-recall for users with coment(u) ≥ 1.
areas, the number of its check-ins and its density. For the
time-related features, we use the number of check-ins at that
day and hour. For the location-related feature, we adopt
location entropy [2] to capture the location’s popularity.
Experiments. We split the dataset from [1] into two: one
from 2011/03/01 to 2011/09/25 for training and the other
from 2011/09/26 to 2011/10/23 testing. We partition the
four metropolises into 0.01×0.01 degree latitude and longitude cells, a user is said to be in a cell if he checked in at the
cell. We construct a balanced dataset for each metropolis.
A support vector machines (SVM) with Gaussian kernel is
exploited as our classifier. For the baseline model, we build
the same set of features out of a user’s all friends.
Result. As shown in Fig. 5, the prediction accuracy is fairly
reasonable (around 70%). With the increase of community
entropy, the accuracy grows faster for the community-based
model which means the predictor works better for users
with high community entropies. This validates our observation that users with high social diversities get more impacts
from their communities than friends. Fig. 6 summarizes the
precision-recall results for users with community entropies
≥ 1. We can conclude that community information can
be explored to achieve promising location predictions, especially for those users with high community entropies.
5.
CONCLUSION AND FUTURE WORK
We have studied influence from communities on user’s mobility (see [3] for details). In the future, we will investigate
community impact on other social behaviors.
6.
REFERENCES
[1] Cho, E., Myers, S. A., and Leskovec, J. Friendship
and mobility: user movement in location-based social
networks. In KDD 2011.
[2] Cranshaw, J., Toch, E., Hone, J., Kittur, A., and
Sadeh, N. Bridging the gap between physical location
and online social networks. In UbiComp 2010.
[3] Pang, J., and Zhang, Y. Location Prediction:
communities speak louder than friends. CoRR
abs/1408.1228, 2014.